Skip to content

Latest commit

 

History

History
34 lines (18 loc) · 1.83 KB

README.md

File metadata and controls

34 lines (18 loc) · 1.83 KB

DeepSynthesis

Implementation and datasets of my master thesis on "LTL Synthesis from Specification Patterns with Neural Networks"

Docker

We recommend to run the data generation and experiments in a Docker container because of the dependencies of this project. More information on Docker on how to install Docker on your machine can be found here. To pull the Docker image from Docker Hub run:

docker pull frederikschmitt/deepsynthesis:latest-cpu

Date Generation

SYNTCOMP

To generate a dataset based on SYNTCOMP patterns run the data_generation_guarantees script and provide the path to the file with the SYNTCOMP patterns and a directory where the dataset is saved:

python -m deepsynthesis.data.synthesis.data_generation.data_generation_guarantees --guarantees /deep-synthesis/data/syntcomp-patterns.json --data-dir /deep-synthesis/data

To list all available options of the data generation run the script with --help.

Grammar

To generate a dataset based on handcrafted specification patterns that were designed using a template grammar run the data_generation_grammar script and provide a directory where the dataset is saved:

python -m deepsynthesis.data.synthesis.data_generation.data_generation_grammar --data-dir /deep-synthesis/data

To list all available options of the data generation run the script with --help.

Training

To train a Transformer on the LTL synthesis problem run the synthesis_transformer_experiment script with a specific dataset and hyperparameters of choice. For example, to train a Transformer on dataset SC100 for 3 epochs run:

python -m deepsynthesis.experiments.synthesis_transformer_experiment --dataset /deep-synthesis/data/SC100 --epochs 3

For an overview of the available hyperparamter options run the script with --help.